arXiv 2601.187532026-03-02cs.LGcs.AI

HalluGuard: 解密数据驱动和推理驱动的LLM幻觉

HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs

CS Department Virginia Tech（弗吉尼亚理工学院计算机科学系）；EECS Department MIT（麻省理工学院电子工程与计算机科学系）；CS Department Dartmouth College（达特茅斯学院计算机科学系）；Statistics Department Virginia Tech（弗吉尼亚理工学院统计学系）；CS Department Michigan State University（密歇根州立大学计算机科学系）

Xinyue Zeng, Junhong Lin, Yujun Yan, Feng Guo, Liang Shi, Jun Wu, Dawei Zhou

备注：: Accepted by The Fourteenth International Conference on Learning Representations (ICLR'26)

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

HalluGuard通过统一理论框架和NTK评分方法，系统识别LLM中的数据驱动和推理驱动幻觉，实现高精度检测。

摘要：

大型语言模型（LLM）在医疗、法律和科学发现等高风险领域中的可靠性常常受到幻觉的损害。这些失败通常源于两个来源：数据驱动的幻觉和推理驱动的幻觉。然而，现有的检测方法通常只解决一个来源，并依赖于任务特定的启发式方法，限制了它们在复杂场景中的泛化能力。为克服这些限制，我们引入了幻觉风险边界，一个统一的理论框架，正式将幻觉风险分解为数据驱动和推理驱动的组成部分，分别与训练时的不匹配和推理时的不稳定性相关。这为分析幻觉如何产生和演变提供了系统的基础。在此基础上，我们引入了HalluGuard，一种基于核诱导度量（NTK）的评分方法，利用NTK诱导的几何和捕获的表示，联合识别数据驱动和推理驱动的幻觉。我们在10个多样化的基准测试、11个竞争性基线和9个流行的LLM基础模型上评估了HalluGuard，一致在检测各种LLM幻觉形式上实现了最先进的性能。我们开源了所提出的模型，地址为https://github.com/Susan571/HalluGuard-ICLR2026。

英文摘要：

The reliability of Large Language Models (LLMs) in high-stakes domains such as healthcare, law, and scientific discovery is often compromised by hallucinations. These failures typically stem from two sources: data-driven hallucinations and reasoning-driven hallucinations. However, existing detection methods usually address only one source and rely on task-specific heuristics, limiting their generalization to complex scenarios. To overcome these limitations, we introduce the Hallucination Risk Bound, a unified theoretical framework that formally decomposes hallucination risk into data-driven and reasoning-driven components, linked respectively to training-time mismatches and inference-time instabilities. This provides a principled foundation for analyzing how hallucinations emerge and evolve. Building on this foundation, we introduce HalluGuard, an NTK-based score that leverages the induced geometry and captured representations of the NTK to jointly identify data-driven and reasoning-driven hallucinations. We evaluate HalluGuard on 10 diverse benchmarks, 11 competitive baselines, and 9 popular LLM backbones, consistently achieving state-of-the-art performance in detecting diverse forms of LLM hallucinations. We open-source our proposed \model{} model at https://github.com/Susan571/HalluGuard-ICLR2026.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。